Available online at www.sciencedirect.

com

Economics Letters 99 (2008) 168 171

www.elsevier.com/locate/econbase

Price deflators and the estimation of the production function

Carmine Ornaghi
Economics Division, School of Social Sciences, University of Southampton, Southampton SO17 1BJ, United Kingdom
Received 1 November 2006; received in revised form 22 May 2007; accepted 22 June 2007
Available online 30 June 2007

Abstract

The use of industry indices to deflate nominal revenues and expenditure in intermediary inputs has been found to lead to lower scale estimates
of the production function. This paper proposes a new approach to solve the estimation biases due to the use of industry deflators which relies on
the use of the firms' labour cost.
2007 Elsevier B.V. All rights reserved.

Keywords: Production function; Price deflators

JEL classification: C23; D24

1. Introduction practice of using industry indices leads to lower scale estimates,

Most of the available firm-level datasets report nominal
revenues and expenditure for materials but they do not report
individual data on output and intermediary input prices. This
data limitation is a major problem for the estimation of pro-
duction functions, given that real output and materials cannot be
computed but only approximated using industry deflators. Kettle
and Griliches (1996) hereafter KG are the first to point out
that the practise of using deflated sales as a proxy for real output
will, ceteris paribus, tend to create a downward bias in the scale
estimate obtained from the production function regressions
(pag. 344). They suggest a solution to determine the true
parameters of the production function which relies on the use of
the aggregate industry output.
In a companion paper, Ornaghi (2006), I use a unique panel
data of manufacturing firms that reports individual prices for
output and intermediary inputs to run two parallel estimates. In
the first nominal revenues and expenditures in materials are
deflated by industry indices (in accordance with standard
practice) while in the second I make use of the information on
firm-level prices. Results obtained in that paper show that the
Tel.: +44 23 8059 2529.
E-mail address: C.Ornaghi@soton.ac.uk.
thus confirming the original KG critique. But contrary to the KG
suggestion, the aggregate industry output does not seem to
provide a viable solution to the problem of deflators as it fails to
consider that there are asymmetric biases in the estimated
coefficients of inputs. Indeed, that paper shows that the use of
industry deflators for expenditure in materials exacerbates the
bias in the estimated labour coefficients (already) introduced by
the use of industry deflators for revenues, but it offsets the bias in
the estimated coefficients for materials.1
This paper defines an alternative approach to the estimation
of a production function that can be used when firm-level prices
are not available. This approach allows obtaining unbiased
estimates of the coefficients of labour and materials and results
consistent with constant return to scale. Two important caveats
to this method deserve mention. First, empirical results show
that the problem of small and statistically insignificant coef-
ficient of capital is invariant to the use of firm-level or aggregate
deflators. The present analysis does not provide any important
1
The importance of using an appropriate deflator for expenditure in intermediary
inputs (and not only for revenues) is confirmed by the results in Mairesse and
Jaumandreu (2005). Using the same data, they define two specifications which
include the same measure of materials and defer only for the measured output, and
they find only minor differences between the two estimations.

0165-1765/$ - see front matter 2007 Elsevier B.V. All rights reserved.
doi:10.1016/j.econlet.2007.06.020
 C. Ornaghi / Economics Letters 99 (2008) 168171
advance towards the understanding of this puzzle. Second, the
method proposed in this paper is applicable to specifications in
differences where any unobservable component that is persis-
tent over time is purged from the residual. The application to
specifications in levels might be problematic, because of
identification problems.2
2. Econometric framework
As in KG, let me assume for simplicity that the production
function of firm i in year t is a CobbDouglas function that can
be written in a log-linear form as:
qit a1 lit a2 mit a3 kit uit
where q, l, m and k are, respectively, the logs of the quantity
produced, labour, materials and capital; u is the random error
term for the equation. The error term u is assumed to include
unobservable firm-specific factors of production, i (e.g.
entrepreneurial ability) that determine persistent productivity
differences between firms over time. Consequently, the error
term can be decomposed as uit = i + vit, where v is a white noise
disturbance term.
Taking logarithms and first differences to eliminate i from
the residual, we obtain the linear equation:
qit a1 lit a2 m it a3 kit vit
where lowercase letters with a tilde stand for logarithmic
differences between year t 1 and t (e.g. q it = lnQit lnQit 1).
When firm-level prices for output Pit and intermediary
inputs Git are not available and the econometrician decides to
use the industry indices PIt and GIt, Eq. (2) is replaced by a new
specification:
yit a1 lit a2 nit a3 kit pit  pIt  a2 git  gIt vit 3 malizing with respect to these industry indices, Eq. (6) can be
eit
where y and refers to the (log differences of) revenues and
expenditure in intermediary inputs deflated by the industry indices,
i.e. y ln PitPTQ it TQit1
 ln Pit1PIt1 and n ln GitGTM it
 ln Git1GIt1
It It
following this transformation, the new disturbance term em-
bodies the deviations of individual output and intermediary price
from the corresponding industry deflators, (p it p It)2(g it g It)
(see Ornaghi, 2006, for details). As the price of output and
intermediary inputs affect the optimal choice of production
factors, the regressors of Eq. (3) are correlated with and this
causes biases in estimating their coefficients. Hereafter, I suggest
a solution to this problem which consists in replacing the (usually)
unknown expression (p it p It) 2(g it g It) with a measure of
labour costs.
2
This is unfortunate since the information contained in levels have been used
to attenuate the bias in the estimation of the capital coefficient, as in the paper
by Blundell and Bond (2000).
169
Assume that firms use the price setting rule Pit = i * MCit,
where MC is the marginal cost and is the constant mark-
up applied by the firm. Taking the log first differences, we
have:
pit ln MCit  ln MCit1 : 4
Solving the cost minimization problem of firm i:
min W TL GTM RTK s:t: Q La1 M a2 K a3
where W, G and R indicate the price of labour, materials and the
rental price of capital, respectively, we obtain the firm cost
1 function:
1 a1 a2 a3
CW ; G; R; Q gTQa1 a2 a3 W a1 a2 a3 Ga1 a2 a3 Ra1 a2 a3 5
where is a constant that depends on the s' coefficients of
the CobbDouglas production function. If returns to scale are
constant (hypothesis supported by our estimates), the
marginal cost can be written as MC = c * W 1 G 2 R 3 . Taking
the first log differences of the latter equation, i.e. ln MCit ln
MCit1 = 1w it + 2 g it + 3r it, and substituting it into Eq. (4), we
obtain that:
pit a2 git a1 w it a3 rit : 6
2
The unknown prices of output and intermediary inputs are
now expressed in terms of the price of labour (generally ob-
servable) and the rental price of capital. Following the approach
used in Klette (1999), this latter equation can be expressed in
terms of deviation from a reference point, which can be thought
of as the price of output and inputs of the representative firm.
Empirically, this reference point is computed as the average
value of the relevant variables within the industry. After nor-
restated as:
pit  pIt  a2 git  gIt a1 w it  wIt a3 rit  rIt : 7
Finally, by replacing Eq. (7) into Eq. (3), we get:
TMit1
;
yit a1 lit a2 nit a3 kit a1 wit  wIt a3 rit  rIt: vit 8
eit
The panel-data used in this study reports all the relevant
information for computing the cost of labour.3 As rental price of
capital depends on the individual rate of return required by a
firm, the price of investment goods and the tax treatment (see
Hall and Jorgenson, 1969, among others), most of which are not
observable, I opt for not computing a proxy for the cost of capital.
The term (r it r It) enters then the error term . It must be considered
that, in the absence of reliable data, the residual would embody the
measurement errors of the rental cost of capital anyway.
3
The value of W for each firm is computed divided the total cost of labour by
the number of employees.
170 C. Ornaghi / Economics Letters 99 (2008) 168171

Table 1 3. Regression results

Production function estimates
I II III Variables are computed using data retrieved from the En-
Eq. (2) a Eq. (3) a Eq. (8) a cuesta sobre Estrategias Empresariales, ESEE, (Business
Labour 0.324 0.187 0.310 Strategy Survey): an unbalanced panel sample of Spanish
(0.068) (0.065) (0.073) manufacturing firms covering the period 19901999. Details
Materials 0.616 0.669 0.615 about the panel-data sample and the variables used for the
(0.047) (0.050) (0.052) estimation can be found in Ornaghi (2006).
Capital 0.043 0.118 0.063
The specifications estimated are those reported in Eqs. (2),
(c.r. to s.) (0.063) (0.061) (0.058)
Wage 0.344 (3) and (8) above. For interpretive reasons, all these equations
(0.150) are restated so that the hypothesis of constant returns can be
Ind Dum Inc. Inc. Inc. tested explicitly (i.e. H0: g = 1 where g 1 + 2 + 3). For
Time Dum Inc. Inc. Inc. instance, Eq. (2) is actually specified as:
Period 199099 199099 199099
N. obs 11,476 11,476 11,476 qit  kit a1 lit  kit a2 mit  kit g  1 kit vit :
Sargan T (df ) 55.4 (55) 54.6 (55) 57.2 (55)
m1 9.54 9.56 9.47 The estimation method is the generalized method of moment
m2 0.25 0.60 0.99 (GMM) applied to panel data. The set of instrumental variables
Estimation method: GMM estimates. Heteroskedasticity robust standard are held fixed across the specification presented below, so that
errors shown in parentheses. Significant at 10% level; significant at 5%; any differences in estimates can be attributed to the change in
significant at 1%.
a
IV s: number of workers, physical capital and materials lagged levels from
the variables used.
t 2 to t 4; (exogenous variables) growth of capital. Table 1 presents the estimated coefficients of the three
different specifications. Column I refers to specification Eq. (2)
in which I use real measures of output and intermediary inputs;
There are two reasons why the bias in estimating the column II corresponds to Eq. (3) based on industry deflators
parameters of specification Eq. (8) can be expected to be smaller while specification in column III refers to Eq. (8), where the term
compared to specification Eq. (3). First, the conditional factor (w it w It) is added to soften the problem created by industry
demands for labour and materials derived from the cost deflators. Results in the first two columns are equivalent to those
minimization problem above respond more to the price of reported in Table 1 of Ornaghi (2006). They show that scale
output and intermediary inputs than to the rental cost of capital. elasticities are lower when deflated sales, Y, and expenditure
Second, to the extend that the rental cost of capital is rather in intermediate inputs, N, are used instead of output, Q, and
constant over time or it has a negligible dispersion within the firms materials, M. In particular, the null hypothesis of constant return
in the industry, the noise term (r it r It) is largely captured by the to scale is rejected at 5% significant level only in column II. The
firm specific effect and industry dummies. Results presented in new interesting finding of this paper is that estimates are notably
the following section seem to confirm this perspective. improved when we use information on labour costs (column III):
Before discussing the empirical results, a final remark is the hypothesis of constant return to scale cannot be rejected and
required. If the assumption of constant markups does not hold, the the estimated elasticity of labour is similar to the value obtained
error term in Eq. (8) would also include an unobserved term in in column I. Moreover, the coefficient of wage is very close to
differences of the markups. Under this alternative scenario, that of labour; the null hypothesis that the two coefficients are
estimated coefficients might be biased because of a possible not statistically different cannot be rejected at any level of
correlation of this unobservable with factor quantities. By com- significance (the p-value of the Wald-test statistic is 0.72), thus
paring the estimates of Eqs. (2) and (8), I find that there are no giving further support to the proposed approach.5
relevant bias due to the omitted markups variable. The question To check whether the differences between the estimated
now is: How can one be sure that estimates are not biased when a coefficients of the variables reported in Table 1 are statistically
similar comparison cannot be done? The econometric framework significant, I compute bootstrap confidence interval by running
defined above suggests that a formal test of the null hypothesis 400 replications of the GMM estimator.6 Table 2 shows the
that the coefficients of labour and wages in Eq. (8) are equal may
5
help to address this question. Given that the correlation of labour As our empirical specifications are defined to test the hypothesis of c.r.s.,
and wages with any unobserved variable (e.g. markups or cost of the well-known problem of small capital coefficient is not immediately
capital) can go in different directions or at least, have different apparent. However, it is easy to see that the capital coefficient takes low values
across all the specifications. As already said, our estimates leave this puzzle
intensities, failing to reject the null hypothesis would provide unsolved. Nevertheless, the results of this paper are not undermined by this
some support in favour of the approach proposed in this paper.4 problem. By defining an alternative measure of capital using data on capacity
utilization, it is possible to get point estimates of the capital coefficient that
are five times higher than those reported in Table 1 (see Ornaghi, 2003, for
4
It is important to note that the existing literature about the cyclical further details). The use of this improved measure of capital does not change
behaviour of markups and real wages is filled with conflicting hypothesis and the finding of this paper with respect to the estimation bias of industry deflators.
inconclusive empirical evidence. This seems to confirm that there are no 6
Note that this estimation are based on the block bootstrapping procedure. This
theoretical reasons or empirical evidence to suspect a bias in the proposed consists in randomly drawing a sample of firms and, for each drawn firm, all the yearly
approach. observations available are used, i.e. the observations of a given firm are kept together.
 C. Ornaghi / Economics Letters 99 (2008) 168171 171

Table 2 Acknowledgement
Test of differences in coefficients
Variables IV V The author would like to thank an anonymous referee for
Eq. (2) vs. Eq. (3) Eq. (2) vs. Eq. (8) useful comments.
Labour 0.117 0.011
[0.065; 0.166] [ 0.094; 0.098] References
Materials 0.044 0.001
[ 0.082; 0.011] [ 0.051; 0.054] Blundell, R., Bond, S., 2000. GMM estimation with persistent panel data: an
Capital 0.065 0.015 application to production functions. Econometric Reviews 19, 321340.
(c.r. to s.) [0.020; 0.106] [ 0.055; 0.072] Hall, R.E., Jorgenson, D.W., 1969. Tax policy and investment behavior. The
Values reported are the bootstrap average differences and, in brackets, the 90% American Economic Review 59, 388401.
Kettle, T.J., Griliches, Z., 1996. The inconsistency of common scale estimators
confidence interval based on 400 replications of the GMM estimations.
when output prices are unobserved and endogenous. Journal of Applied
Econometrics 11, 343361.
Klette, T.J., 1999. Market power, scale economies and productivity: estimates
from a panel of establishment data. The Journal of Industrial Economics 48,
average differences of the estimated coefficients together with 451476.
90% confidence interval in square brackets. Column IV reports Mairesse, J., Jaumandreu, J., 2005. Panel-data estimates of the production function
the difference between Eqs. (2) and (3) (as in Table 2 of Ornaghi, and the revenue function: what differences does it makes? Scandinavian
2006) while column V those between Eqs. (2) and (8). The Journal of Economics 107, 651672.
confidence intervals confirm that there are not significant dif- Ornaghi, C., 2003. Assessing the effects of measurement errors on the estimation
of the production function. Working Papers 02-33, Universidad Carlos III de
ferences between estimation based on firm-level deflators and Madrid.
on industry deflators when the wage term is included in the Ornaghi, C., 2006. Assessing the effects of measurement errors on the estimation
specification. of production functions. Journal of Applied Econometrics 21, 879891.